Research note
Why I built Principles of Earned Autonomy
An agent can finish a task without making its work easy to take responsibility for. That gap is the starting point.
Completion is not the whole problem
When an AI agent delivers a plausible result, the person responsible still needs to know what it understood, which decisions changed the direction and what was actually verified. A confident completion message answers none of those questions by itself. My interest is in the architecture of trustworthy delegation: how a person can hand over real work while retaining meaningful authority and grounds for examining the outcome. Principles of Earned Autonomy is my proposed response, not an external certification that the problem is solved.
Three connected principles
Operator’s Intent separates the destination from a prescribed route. The human supplies purpose and consequential boundaries; the agent must interpret and adapt. Observable Autonomy requires inspectable evidence of autonomous work rather than invisible action. Convergence Is Silence makes independent challenge part of deciding when to stop. The framework’s operational criterion calls for fresh evaluations from three distinct model families, not repeated self-approval in one conversation. Even that stopping signal is not a guarantee of correctness or safety.
From framework to working practice
The PEA Skills Suite is a reference implementation in developer tooling. Improve is its normal entry point, supported by intent interpretation, persistent context, a decision trail and reflection. That makes the framework and the suite connected but distinct: one proposes architectural principles; the other asks agents to apply a workflow. Markdown instructions do not enforce themselves. A skill appearing in an installation or registry does not establish that an agent read it, followed it or produced useful evidence. The first public pilot makes that distinction concrete.
The contribution must remain challengeable
The work is open so people can inspect the argument, run the tools and identify where they fail. An agent-authored explanation is not proof of faithful internal reasoning, and a useful product is not a controlled demonstration that PEA caused its delivery speed. My aim is to connect human direction, observable work and retained context in a way that practitioners can test. Independent attempts, negative results and informed disagreement matter more than an impressive label. This is my independent research and engineering work, separate from my employment.